What’s new with

Andrew DavisonUNIC, CNRS

FACETS CodeJam #2Gif sur Yvette, 5th-8th May 2008

Outline

1 A brief introduction to PyNN

2 A tour of the API

3 Parallel simulations

4 Use cases

5 Future directions

Simulator diversityProblem and opportunity

Cons

Considerable di!culty in translating models from one simulator toanother...

...or even in understanding someone else’s code.

This:

impedes communication between investigators,makes it harder to reproduce other people’s work,makes it harder to build on other people’s work.

Pros

Each simulator has a di"erent balance between e!ciency, flexibility,scalability and user-friendliness ! can choose the most appropriatefor a given problem.

Any given simulator is likely to have bugs and hidden assumptions,which will be revealed by cross-checking results between di"erentsimulators ! greater confidence in correctness of results.

Simulator-independent model specification(“Meta-simulators”)

Simulator-independent environments for developing neurosciencemodels:

keep the advantages of having multiple simulators

but remove the translation barrier.

Three (complementary) approaches:

GUI (e.g. neuroConstruct)

XML-based language (e.g. NeuroML)

interpreted language (e.g. Python)

A common scripting language for neuroscience simulators

Simulator LanguagePCSIM C++ or PythonMOOSE SLI or PythonMVASpike C++ or PythonNEST sli or PythonNEURON hoc or PythonSPLIT C++ (Python interface planned)Brian PythonFACETS hardware Python

A common scripting language for neuroscience simulators

Goal

Write the code for a model simulation once, run iton any supported simulator* without modification.

* or hardware device

Architecture

sli

GENESIS 2 GENESIS 3(MOOSE)

NeuroML

PCSIMNEST NEURONSimulator kernel

Native interpreter

Python interpreter

Simulator-specific PyNN module

hoc

FACETShardware

nrnpy

SLI

PyGENESISPyPCSIM PyHALPyNEST

pynn.neuronpynn.nest pynn.pcsim pynn.facetshardware1

pynn.neuroml pynn.genesis3

pynn.genesis2

PyNN

Direct communication Code generation Implemented Planned

How to get PyNN

Latest stable versionhttp://neuralensemble.org/PyNN/wiki/Download

Latest development versionsvn co https://neuralensemble.org/svn/PyNN/trunk pyNN

Full documentationhttp://neuralensemble.org/PyNN

Installing PyNN

via svn or distutils

How to participate in PyNN development

http://neuralensemble.org/PyNN

How to participate in PyNN development

Google groups screenshot

Google groups URL

Outline

1 A brief introduction to PyNN

2 A tour of the API

3 Parallel simulations

4 Use cases

5 Future directions

Selecting the simulator

from pyNN.neuron import *from pyNN.nest1 import *from pyNN.nest2 import *from pyNN.pcsim import *from pyNN.moose import *from pyNN.brian import *

import pyNN.neuron as sim

setup() and end()

setup(timestep=0.1, min_delay=0.1, debug=False)

setup(timestep=0.1, min_delay=0.1, debug=’pyNN.log’,threads=2, shark_teeth=999)

end()

create()

create(IF_curr_alpha)

create(IF_curr_alpha, n=10)

create(IF_curr_alpha, {’tau_m’: 15.0, ’cm’: 0.9}, n=10)

>>> IF_curr_alpha.default_parameters{’tau_refrac’: 0.0, ’tau_m’: 20.0, ’i_offset’: 0.0,’cm’: 1.0, ’v_init’: -65.0,’v_thresh’: -50.0,’tau_syn_E’: 0.5, ’v_rest’: -65.0, ’tau_syn_I’: 0.5,’v_reset’: -65.0}

create()

>>> create(IF_curr_alpha, param_dict=’foo’: 15.0)Traceback (most recent call last):

.

.

.NonExistentParameterError: foo

>>> create(IF_curr_alpha, param_dict=’tau_m’: ’bar’)Traceback (most recent call last):

.

.

.InvalidParameterValueError:

(<type ’str’>, should be <type ’float’>)

create()

create(IF_curr_alpha, ’v_thresh’: -50, ’cm’: 0.9)

create(’iaf_neuron’, ’V_th’: -50, ’C_m’: 900.0)

Standard cell models

IF_curr_alphaIF_curr_expIF_cond_alphaIF_cond_exp,IF_cond_exp_gsfa_grrIF_facets_hardware1HH_cond_exp,EIF_cond_alpha_isfa_istaSpikeSourcePoissonSpikeSourceInhGammaSpikeSourceArray

Standard cell models

Example: Leaky integrate-and-fire model with fixed firingthreshold, and current-based, alpha-function synapses.

Name Units NEST NEURONv rest mV U0 v restv reset mV Vreset v resetcm nF C† CMtau m ms Tau tau mtau refrac ms TauR t refractau syn ms TauSyn tau synv thresh mV Theta v threshi offset nA I0† i offset

†Unit di"erences: C is in pF, I0 in pA.

ID objects

>>> my_cell = create(IF_cond_exp)>>> print my_cell1>>> type(my_cell)<class ’pyNN.nest2.ID’>>>> my_cell.tau_m20.0>>> my_cell.position(1.0, 0.0, 0.0)>>> my_cell.position = (0.76, 0.54, 0.32)

connect()

spike_source = create(SpikeSourceArray,{’spike_times’: [10.0, 20.0, 30.0]})

cell_list = create(IF_curr_exp, n=10)

connect(spike_source, cell_list)

connect(sources, targets, weight=1.5, delay=0.5,p=0.2, synapse_type=’inhibitory’)

record()

record(cell, "spikes.dat")

record_v(cell_list, "Vm.dat")

Writing occurs on end()

run()

run(100.0)

Simulation status

get_current_time()

get_time_step()

get_min_delay()

num_processes()

rank()

Random numbers

>>> from pyNN.random import NumpyRNG, GSLRNG, NativeRNG

>>> rng = NumpyRNG(seed=12345)>>> rng.next()0.6754034>>> rng.next(3, ’uniform’, (-70,-65))[-67.4326, -69.9223, -65.4566]

Use NativeRNG or GSLRNG to ensure di"erent simulators getthe same random numbers

Use NativeRNG to use a simulator’s built-in RNG

Random numbers

>>> from pyNN.random import RandomDistribution

>>> distr = RandomDistribution(’uniform’, (-70, -65),... rng=rng)>>> distr.next(3)[-67.4326, -69.9223, -65.4566]

Populations

p1 = Population((10,10), IF_curr_exp)

p2 = Population(100, SpikeSourceArray,label="Input Population")

p3 = Population(dims=(3,4,5), cellclass=IF_cond_alpha,cellparams={’v_thresh’: -55.0},label="Column 1")

p4 = Population(20, ’iaf_neuron’, {’Tau’: 15.0,’C’: 100.0})

PopulationsAccessing individual members

>>> p1[0,0]1>>> p1[9,9]100>>> p3[2,1,0]246

>>> p3.locate(246)(2, 1, 0)

>>> p1.index(99)100

>>> p1[0,0].tau_m = 12.3

PopulationsIterators

>>> for id in p1:... print id, id.tau_m...0 12.31 20.02 20.0...

>>> for addr in p1.addresses():... print addr...(0, 0)(0, 1)(0, 2)...(0, 9)(1, 0)(1, 1)(1, 2)...

set(), tset(), rset()

>>> p1.set("tau_m", 20.0)

>>> p1.set(’tau_m’:20, ’v_rest’:-65)

>>> distr = RandomDistribution(’uniform’, [-70,-55])>>> p1.rset(’v_init’, distr)

>>> import numpy>>> current_input = numpy.zeros(p1.dim)>>> current_input[:,0] = 0.1>>> p1.tset(’i_offset’, current_input)

Recording

# record from all neurons in the population>>> p1.record()

# record from 10 neurons chosen at random>>> p1.record(10)

# record from specific neurons>>> p1.record([p1[0,0], p1[0,1], p1[0,2]])

>>> p1.printSpikes("spikefile.dat")

>>> p1.getSpikes()array([])

Position in space

>>> p1[1,0].position = (0.0, 0.1, 0.2)>>> p1[1,0].positionarray([ 0. , 0.1, 0.2])

>>> p1.positionsarray([[...]])

>>> p1.nearest((4.5, 7.8, 3.3))48>>> p1[p1.locate(48)].positionarray([ 4., 8., 0.])

Projections

prj2_1 = Projection(p2, p1, AllToAllConnector())

prj1_2 = Projection(p1, p2, FixedProbabilityConnector(0.02),target=’inhibitory’, label=’foo’,rng=NumpyRNG())

Connectors

AllToAllConnector

OneToOneConnector

FixedProbabilityConnector

DistanceDependentProbabilityConnector

FixedNumberPostConnector

FixedNumberPostConnector

FromFileConnector*

FromListConnector

(* cf Projection.saveConnections(filename))

Connectors

c = DistanceDependentProbabilityConnector("exp(-abs(d))",axes=’xy’,periodic_boundaries=(500, 500, 0),weights=0.7,delays=RandomDistribution(’gamma’, [1,0.1])

)

Weights and delays

>>> prj1_1.setWeights(0.2)

>>> weight_list = 0.1*numpy.ones(len(prj2_1))>>> weight_list[0:5] = 0.2>>> prj2_1.setWeights(weight_list)

>>> prj1_1.randomizeWeights(weight_distr)

>>> prj1_2.setDelays(’exp(-d/50.0)+0.1’)

[Note: synaptic weights are in nA for current-based synapses andµS for conductance-based synapses]

Weights and delays

w_array = prj.getWeights()

prj.printWeights(filename)

Synaptic plasticity

# Facilitating/depressing synapsesdepressing_syn = SynapseDynamics(

fast=TsodyksMarkramMechanism(**params))prj = Projection(pre, post, AllToAllConnector(),

synapse_dynamics=depressing_syn)

# STDPstdp_model = STDPMechanism(

timing_dependence=SpikePairRule(tau_plus=20.0,tau_minus=20.0),

weight_dependence=AdditiveWeightDependence(w_min=0, w_max=0.02,A_plus=0.01, A_minus=0.012)

)prj2 = Projection(pre, post, FixedProbabilityConnector(p=0.1),

synapse_dynamics=SynapseDynamics(slow=stdp_model))

Outline

1 A brief introduction to PyNN

2 A tour of the API

3 Parallel simulations

4 Use cases

5 Future directions

mpirun

png from NEURON on di"erent numbers of processors?

Outline

1 A brief introduction to PyNN

2 A tour of the API

3 Parallel simulations

4 Use cases

5 Future directions

Use cases

Testing a model on multiple simulatorsCross-checking gives greater confidence in correctness of results.

Porting a model between simulatorsGradually replace simulator-specific code with Python code,checking results are unchanged at each step.

Hardware interfaceNeuromorphic VLSI hardware can also use PyNN, allowing directcomparison of numerical simulations and emulations in silicon.

Collaborating between di!erent groupsEach group can use their preferred simulator, while working on acommon code base.

Outline

1 A brief introduction to PyNN

2 A tour of the API

3 Parallel simulations

4 Use cases

5 Future directions

Future directions

Extend range of simulators supported (full NeuroML support,MOOSE, Brian,...)

Support non-spiking (firing-rate based) neuron models?

Support explicit units (cf Brian)

Optimisation, so PyNN is only a little slower than native code

Improved parallelisation

Extensions of the API:current highest level of organisation is Population, Projection.extend to Column, Layer, Meta-column, Map, ...extend stimuli, e.g., DriftingGrating, DenseNoise,...extend recording, e.g. recordActivityMap(), ...